Information Technology Reference
In-Depth Information
Fig. 4.9
Topic-sensitive influencer identification performance comparison
The results are shown in Fig.
4.9
. We can see that the proposed mmTIM consis-
tently outperforms the two baselines. Note that the average #contact user is around
15. The top-1 accuracy of mmTIM is about 25%, which indicates that one out of
four trials, mmTIM succeeds to identify the real topic-sensitive influencer at the first
rank. The problem of TAP can be summarized in twofold: First, TAP assumes that
topic space and node topic distribution are available before social influence mod-
eling. However, the predefined topic space extracted from user-annotation may be
insufficient to capture the semantics in social influence links. Second, whenmodeling
topic-sensitive influence, TAP only utilizes the link information of user-contact user,
which loses information of relations between users and images. mTIH obtains better
performance than TAP by addressing the above two problems. However, mTIH is
proposed for text-based citation networks and does not address the multimodality
issue, which results in the inferior performance to mmTIM. Moreover, compared
with mmTIM, mTIH is more focused on the document-level (i.e., image or chapter)
instead of user-level, by explicitly building the influencer set from user-document
relations. While, under social media settings, the influencer set of specific user is
actually available by follow or friend list.
4.5.3 Personalized Image Search Evaluation
The derived topic-sensitive influence values can be applied to applications like social
search, friend recommendation, group suggestion, etc. Based on the proposed risk
minimization-based theoretical framework, we evaluate the effectiveness of mmTIM
on personalized image search and topic-based image recommendation.
In the query and image language models, according to Eq. (
4.16
), there is a weight
parameter
based on
the assumption that the searcher himself/herself should be more trustworthy if he/she
has much annotation activities, otherwise the influencer should be more trusted.
Formally, for searcher
U
i
,
ˁ
controlling the strength of searcher and influencer. We choose
ˁ
ˁ
is set as:
|
T
U
i
|
ˁ
=
U
j
∈
C
U
i
|
T
U
j
|
1
|
C
U
i
|
|
T
U
i
|+
